COREGISTRATION BASED ON SIFT ALGORITHM FOR SYNTHETIC APERTURE
RADAR INTERFEROMETRY
Fangting Li a, *, Guo Zhang a, Jun Yan a
a
State key Library of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University,
Wuhan, 430079, China, lifangting1985@163.com, guozhang@whu.edu.cn, yanjun_pla@263.net
KEY WORDS: Coregistration , INSAR, SIFT,
ABSTRACT:
Single-look complex image coregistration is the key step of Synthetic Aperture Radar(SAR) Interferometry. The precision of the
coregistration results have a direct effect on the quality of the SAR interferogram generated, thereby it will influence the accuracy of
extracting DEM. In this paper, SIFT (Scale Invariant Feature Transform) algorithm will be used in Single-look complex image
coregistration, and in the pilot study, experiments have provn the method is useful.
purpose of coarse matching is to provide an initial value for the
stereo-pair pixel searching of precise matching.
1. INTRODUCTION
1.1 Single-look complex image coregistration
The precise matching method of SLC images, firstly, samples
the master image and slave image, then identifies N uniform
distributed control feature points on the master image, and
selects the matching window of a certain size, whose center is
the control feature points. According to the deviation values of
coarse matching, the precise matching method then selects a
larger search window than the matching window in the
corresponding position of the slave image. Based on a certain
order, it moves the matching window pixel by pixel to calculate
the indicator values for the two windows. The point with the
best matching indicator value in the search window will be
chosen as the stereo-pair point in the slave image. Through
these processes, the coordinates in two images of the stereo-pair
points are obtained.
Single-look complex (SLC) image coregistration is the key step
of synthetic aperture radar interferometry. The precision of the
coregistration results have a direct effect on the quality of the
SAR interferogram generation, thereby it will influence the
accuracy of extracting DEM. When accurately coregistrater the
two SAR complex images from the two approaching tracks,
their interferometric phase differencing images will display
stripes. The changes of the stripes include the terrain undulation
information. If the two images were not precisely
coregistratered, the interference stripes generated will be
blurred, or even no interference stripes will be generated. At
present, in order to achieve sub-pixel level accuracy, the
generally used coregistration method for complex image is
multi-stage coregistration, the commonly used method is: the
rough image coregistration based on the orbit information or on
the intervention of users, which is the first stage of
coregistration; image coregistration based on pixel level, which
is the second stage of coregistration; and image coregistration
based on sub-pixel level, which is the third stage of
coregistration.
1.2 Original solutions
SLC image coregistration is to calculate the coordinate
projection relationship between a master image and slave image,
and then to take use of this relationship to implement
coordinates transformation, image interpolation and resampling.
The interference measurement requires the accuracy of image
matching to be on a sub-pixel level. Therefore, the SAR image
matching includes two steps - coarse and precise matching.
The coarse matching can be effectuated by using satellite
orbital parameters or manually selecting a few feature points to
calculate the deviation values, Δr and Δc, in direction (row of
image) and in distance (column of image) between the master
image and slave image. The deviation values are relatively
rough values, whose accuracy is usually on pixel level. The
According to the coarse and precise matching mentioned above,
we can get N coordinate pairs of N feature stereo-pair points.
The polynomial (such as third-order) model is used to simulate
the coordinate projection relationship of the master image and
slave image. The parameters of the polynomial model can be
solved by measurements of N coordinate pairs and the leastsquares algorithm. The coordinate transformation relationship
of the pair-image is achieved. Finally, coordinates
transformation and resampling can be carried out based on this
relationship. In this way, the slave image is transformed into the
space of master image, which is ready to produce the
interferometry image.
The three matching method analyzed in this study are all
precise matching methods. The general algorithms and the
processes are all almost the same with these three methods. The
only differences are the method for calculating the indicator
value in the matching window and the standard of selecting the
stereo-pair point. The details of these three methods will be
discussed below.
(1) The coherent coefficient method uses the coherent
coefficient as the indicator value. The coherent coefficient of
the stereo-pair pixel of matching window and the target window
* Corresponding author.
123
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008
image data that has been transformed relative to the assigned
orientation, scale, and location for each feature, thereby
providing invariance to these transformations.
Then, the local image gradients are measured at the selected
scale in a region around each keypoint. These are transformed
into a representation that allows for significant levels of local
shape distortion and change in illumination (Lowe, 2004).
can be calculated by formula (1). This coherent coefficient is
the value of the window centre.
m
n
i =1
i =1
∑ ∑ M (i, j )S
ϒ=
m
n
i =1
i =1
*
(i, j )
m
n
i =1
i =1
∑ ∑ M (i, j ) ∑ ∑ S (i, j )
2
(1)
2
2.2 Advantages of the SIFT algorithm
Theoretically speaking, the SIFT algorithm is invariant, even
for images with scale change and rotation. However, the
tectonic of SIFT has been specially treated in many details.
Therefore, SIFT algorithm has a strong adaptation to images
with complex deformation and changes of light. At the same
time, it has higher computing speed and higher positioning
accuracy.
In the formula, M (i, j), S (i, j) are the plural data of the corresponding
position (i, j) in two matching windows respectively. The symbol “*”
represents conjugated plural. The coherent coefficient of every
point in the search window can be calculated by the formula
above. The point with the largest coherent coefficient is
selected as the matching point.
(1) Compared to the traditional method Laplacian of Gaussian
(LoG), DoG has higher computing speeds to detect the keypoint
in scale space.
(2) The precise position of the keypoint not only improves the
accuracy, but also improves the stability of the keypoint.
(3)When constructing the keypoint descriptor, we use statistical
characteristics on a sub-region level as a research object, not on
a pixel level, which improves the adaptability to the local
deformation of images (Zhao, et al., 2007).
(2) The greatest spectrum method uses the spectrum value as
the indicator value. The two matching windows are processed
by interference processing to obtain the interference fringe
images. The two-dimensional discrete Fourier Transformation
(DFT) is processed to the interference fringe images to get the
two-dimensional spectra. The greatest value of the plural
absolute value of the two-dimensional spectra is the indicator
value (spectra value). Finally, the point with the greatest spectra
value is selected as the matching point.
2.3 Coregistration based on sift for insar
(3) The average fluctuation function of the phase difference
method uses the function value (f) of the phase difference
average fluctuation as the indicator value. Firstly, the
corresponding pixel phase differences P (i, j) of the two
matching windows are calculated, and the function value (f) of
the phase difference average fluctuation is calculated by
formula (2).
f =∑
i
∑ ( P(i + 1, j) − P(i, j) + P(i, j + 1) − P(i, j) ) / 2
(2)
j
Where,f is the indicator value. The point with the least f value is
selected as the matching point (Lu, 2005).
2. METHODOLOGIES
2.1 SIFT algorithm
2.1.1 Keypoint detection
The first step of SIFT construction is the detection of a keypoint.
The main principium: for each octave of scale space, the input
image is convolved with the Gaussian function to produce the
set of scale space images. And then adjacent Gaussian images
are subtracted to produce the difference-of-Gaussian images
(DoG). Finally, the Gaussian image is Sub-Sampled by a factor
of 2. A pixel is compared to its 26 neighbors in 3 by 3 regions
at the current and adjacent scales, detecting the maxima and
minima of the difference-of-Gaussian images.
In addition, with the curve fitting method, the keypoint can be
further processed by precise location.
Some commonly used operators for describing characteristics
are Sum of Squared Difference (SSD), Sum of Absolute
Difference (SAD), and Normalized Cross Correlation (NCC).
Directly depending on gray information of images, all these
operators are sensitive to noise in the images. Thus, the
robustnesses of these operators are weak during the non-linear
gray transformation of images. As for a SAR image with mass
speckle noise, using these operators seems to be unpractical.
However, the method based on SIFT algorithm shows a
characteristic of better robustness and anti-interference when
transforming images in both geometric and optical aspects. On
the basis of this conclusion, the SIFT operator can be applied to
the registration of INSAR image processing for getting a better
result. This result can also be used for further steps of
interference processing. In this paper, the SIFT algorithm will
be used in precise matching to improve the accuracy of
matching and computing speed. The process of coregistration
based on SIFT algorithm is as follows in Fig. 1.
Master
SIFT detect
local image
Slave
SIFT detect
local image
Keypoint match
2.1.2 The local image descriptor
Before the local image descriptor, one or more orientations are
assigned to each keypoint location based on local image
gradient directions. All future operations are performed on
Rectification and coregistration
Fig.1. the process of coregistration
124
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008
then a threshold, we choose the keypoint as a matching point.
The smaller the threshold is, the less matching points we get,
and the more reliable the matching points are (Lowe, 2004).
2.4 Keypoint detection in scale-space
Images can be expressed with different scales, and with the
increasing of the scale parameter, images get smaller. In other
words, a small scale is corresponds to details of image features
and a large scale is for profiling characteristics.
2.7 Rectification and Registration based on TIN
In this part, we create the Triangulated Irregular Network (TIN)
by the minimum distance method (Li, et al., 2006) in the two
images. Each of the large numbers of triangles has three tie
points Xi ,Yi ) ,( X′i ,Y′i ) ,i = 1 ,2 ,3 ,which can be used to
calculate affine parameters:
The scale-space image is defined as the convolution of a
variable-scale Gaussian with an input image
L ( x, y , δ ) = G ( x, y , δ ) * I ( x, y )
(3)
where (x,y) are pixel coordinates of the image and
δ
X′= a0 + a1 X + a2 Y
Y′ = b0 + b1 X + b2 Y
is scale factor.
DoG is computed from the difference of two nearby scales
separated by the constant multiplicative factor k:
D ( x, y, δ ) = (G ( x, y, kδ ) − G ( x, y, δ )) * I ( x, y )
(7)
Three points can get six equations, then we can calculate
a0 ,a1 ,a2 and b0 ,b1 ,b2 with which we can correct theΔP′1
P′2 P′3 on the slave image to ΔP1 P2 P3 on the master
image. The process of rectification is as follow in Fig. 2. (Liu,
et al., 2007)
(4)
The DoG function is similar to the scale-normalized LoG.
Create a buffer for registration image
Every keypoint has information including location, gradient
magnitude and orientation. The scale of the keypoint is used to
select the Gaussian smoothed image, L, with the closest scale,
so that all computations are performed in a scale-invariant
manner. For each image sample, L(x, y), at this scale, the
gradient magnitude, m(x, y), and θ ( x, y ) , is precomputed using
pixel differences:
m(x, y) = (L(x +1, y) − L(x −1, y))2 + (L(x, y +1) − L(x, y −1))2
Get the pixels in the buffer one by one
Decide which triangle the pixel belong
Calculate the corresponding pixel in the master image
(5)
Interpolation
θ(x, y) = tan ((L(x, y +1) − L(x, y −1))/(L(x +1, y) − L(x −1, y))) (6)
−1
No
Is the last pixels
2.5 Produce local image descriptor
Yes
Over
A keypoint descriptor is created by first computing the gradient
magnitude and orientation at each image sample point in a
region around the keypoint location. These are weighted by a
Gaussian window, indicated by the overlaid circle. These
samples are then accumulated into orientation histograms
summarizing the content over 4 by 4 subregions, as shown in
Fig. 2, with the length of each arrow corresponding to the sum
of the gradient magnitudes near that direction within the region.
Fig. 2 shows a 2 by 2 descriptor array computed from an 8 by 8
set of samples.
Fig.2. the process of resampling
3. RESULTS
In this paper, two ERS images of Tianjing（the obtaining time
of the images are 199710190253，199710180253） are used,
one for master image and other for slave image, to test the
method. SIFT algorithm was used to detect the keypoints in the
images which is shown on Fig.3.
2.6 Keypoint match
The best candidate match for each keypoint is found by
identifiying its nearest neighbour in the database of keypoints
from the master image and slave image. The nearest neighbour
is defined as the keypoint with the minimum Euclidean distance
for the invariant descriptor vector.
These keypoints are matched with the method above, which is
shown on Fig.5. Fig.4 illustrates the matdhing points with the
threshold of 0.75. Finally, the slave image is rectified based on
TIN. The two point linked by a white line are the matching
points.
When comparing the distance of the closest neighbour to that of
the second-closest neighbour, if the ratio between them is less
125
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008
keypoints on master image
keypoints on slave image
Fig.3. keypoints
matching points on master image
matching points on slave image
Fig. 4. matching points on two images
126
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008
Fig.4. Keypoint match
advantages when using this method are a higher speed of
computation, a higher precise of matching and an anti-noise
ability.
4. DISCUSSION
In this paper, the SIFT algorithm is introduced into the
coregistration for the INSAR, and the result of our experiments
shows that this method can precisely detect the keypionts and
choose some reliable matching points. Comparing with
traditional method of coregistration for INSAR in frequency
domain, this method has a higher speed.
From the experiments above, we get a prospective result of
coregistration in INSAR process.
REFERENCES
Fig.4 illustrates that there are also few wrong matching points.
We will do some research on
Zhao, W., Xu, W. and Yu, Y., 2007. Area Harmony
Dominating Rectification Method for SIFT Image Matehing.
The Eighth Internayional Conference on Electronic
Measurement and Instruments, ICEMI, vol.2, pp. 935-939.
5. CONCLUSION
In this paper, we have presented a method based on the SIFT
algorithm to solve the problem of INSAR. Coregistration is a
key step in the INSAR process, and a high accuracy of
coregistration has a vital effect on interference. Some
Zhang, C., Gong, Z. and Sun, L., 2008. The modified algorithm
of image matching based on SIFT. Computer Engineering and
Applications, 44(2), pp. 95-97.
127
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B1. Beijing 2008
LiuBin, ZhangGuo, ZhuXiaoyong, GongJianya. “A Method on
High-Precision Rectification and Registration Multi-Source
Remote Sensing Imagery”. Isprs.2007.
Zai, X. and Zhao,Y., 2007. A Stereo Matching Algorithm Based
on SIFT Feature Descriptor. Micor-computer Information, 23
(8-3), pp. 285-287
Zhang, C., Zhou, Y., Wu, S. and Lin, H., 2008. Automatic
mosaic of surveillance images based on SIFT feature matching.
Computer Applications, 28(1), pp. 191-194.
Qian, S. and Zhu, J., 2007. Improved SIFT-based Bidirectional
Image Matching Algorithm. Mechanical Science and
Technology for Aerospace Engineering, 26(9), pp. 1179-1182.
LuLijun. ”Redistration of INSAR image and parallel Algorithm
Study”. 2005.
David. G. Lowe, “Distinctive ImageFeatures from ScaleInvariant Keypoint”. January 5, 2004.
David G. Lowe, Object Recognition from Local Scale-Invariant
Features. Proc. of the International Conference on Computer
Vision, Corfu, 1999.
Li, X., Zhang, L. and Hu, Z., 2006. SIFH Based Automatic
Registration of Remotely-sensed Imagery. Journal of Remote
Senseing, 10(6), pp. 885-892.
ACKNOMLEDGEMENT
The authors wish to thank open research subject of Key
Laboratory of Geo-informatics of State Bureau of Surveying
and Mapping (Project No.: A1721), and China International
Science and Technology Cooperation Project: High-Resolution
Stereo Mapping Satellite: Field Geometric Calibration and
Application (Project No.: 2006DFA71570), and Commission of
Science Technology and Industry for National Defense Project:
Key Techniques of Data Processing for Mapping Satellite for
financial support.
128